Hi, I'm @avcton Muhammad Ahmad
Building production AI systems that ship. From agentic apps to full-stack products.
About
It started earlier than most.
My journey in tech started earlier than most. I was ten when I installed my first Ubuntu, in seventh grade when I shipped my first Android app, in tenth when I built a Hackintosh just to see if I could. The tools changed. The pull didn't.
Now I build production AI at Dubizzle Labs; agentic systems, orchestration infra, and voice agents for Bayut, Zameen & dubizzle across the GCC. Alongside it, an MS in AI at LUMS.
Now
What I'm working on.
Building and running a multi-pod n8n orchestration platform on Kubernetes. It connects Dubizzle Group's products to its AI services (voice agents, WhatsApp, CRM, and client memory) across Saudi Arabia, UAE, and Pakistan, dispatching tens of thousands of AI agent runs a day.
Workflow executions
450k+
Peak daily volume
~23.4k
AI calls placed
100k+
Call conversion
6.2%
Selected work
Things I've shipped.
Spotlight
Voice AI @ Dubizzle
Production calling agents that speak natively: Arabic for Bayut KSA, Urdu for Zameen PK. 6.2% conversion against 1.4% for human telesales.
Interview Sensei
A living AI avatar that interacts & interviews you in real-time. 2× National Competitions Winner
More builds
- Genomaly GAN/VAE anomaly detection applied to pharma quality control. 82% accuracy on capsule defects. Computer Vision
- Visionary GenAI ViT for spoof detection, CLIP for text-image retrieval over 5K images, Stable Diffusion for generation. Three tasks, one pipeline. Generative AI
- Snake AI Side-by-side A*, GBFS, and UCS agents in a live multi-processing snake game built in Python. Search Agents
- Multi-Label Genre Prediction Predicts up to 27 movie genres from a plot synopsis using a two-layer BiLSTM. Deployed live on Streamlit. NLP
- NS3 Docker Utility Pre-configured NS3 network simulation environment on Docker. Adopted by an entire CS section for mid-exams; ~500 organic visitors. DevOps
- Fiscal Count .NET ERP for a real manufacturing client. Led a team of 5 engineers. From prototype to first production system shipped. Full-Stack
Experience
Track.
-
ML Engineer - I
- Architected and built the group's self-hosted n8n orchestration platform solo on Kubernetes (AWS EKS): 450k+ workflow executions since the Jun 2026 go-live, peaking at ~23.4k a day.
- Designed its Fastify auth and throttling gateway end to end: from 90% HTTP 503s at 5,000 users to 100% success on 1,056 concurrent requests at 377 req/s.
- Automated retries and campaigns with no human in the loop: ~21.8k leads under retry handling and 48 campaigns self-provisioned with zero manual DB work.
- Led agentic framework R&D across six platforms; a live multi-agent demo to the CEO in a 30+ person boardroom was approved company-wide.
Associate ML Engineer
- Built and led Zara, an Arabic AI calling agent for Bayut KSA: 6.2% conversion vs. 1.4% and 0.8% for human telesales, with results in 1.5 days instead of ~2 weeks.
- Diagnosed seven categories of production failure, then engineered a dialect-native translation layer and led voice cloning end to end.
- Trained a 13-head Longformer to flag hallucinations per field in LLM call analysis; an LSTM over Whisper confidence scores reached 83% accuracy.
- Opened a telephony carrier relationship through cold outreach, stood up SIP trunking for GCC calling, and deployed Dubizzle's first production n8n.
-
- Built Interview Sensei end-to-end: real-time AI mock interviews with a 3D avatar, WebRTC video/audio, STT/TTS, LLM-driven question generation, and post-session analytics.
- Deployed on Nextbridge's RTX 4090 via NGINX reverse proxy with GPU session rate-limiting for concurrent interview sessions.
- 1st at FAST Job Fair 2025 (250+ FYPs; first-ever DS team to win) and 1st at SOFTEC 2025 SPC.
-
- Built a Flutter mobile app with custom theming, routing, and session management; backed by a Dockerized Node.js/Express + MongoDB service.
- Optimized DB query execution by 30% via Mongoose ORM; delivered 20+ API endpoints across 10+ sprint cycles over an 8-month remote engagement.
-
- Engineered a PHP CodeIgniter data model processing 50,000+ airline ticket records monthly; improved data handling efficiency by 30%.
- Automated XML airline data parsing into target schema; cut manual processing time by ~40% at ~99% accuracy.
Stack & Tools
What I build with.
Education
School.
Studying alongside full-time work.
FYP: Interview Sensei. First DS team to win Job Fair + SOFTEC SPC 2025.
Recognition
Receipts.
Voices
In their words.
-
Rare combination of research depth and practical engineering. Interview Sensei was ready to ship the day we judged it.
Amir Quddus Industry Judge SOFTEC 2025 -
One of the sharpest students I've taught. His final project won two national competitions; the engineering quality was production-ready from day one.
Anonymous Professor Faculty FAST NUCES
Digital Garden
What I've been thinking about.
Notes, essays, half-finished thoughts. Tended in public.
How We'd Work Together
Got something we can build? Here's the journey.
From first conversation to a working system in production. Five steps · ~6 weeks to v1.
- 01 INTRO
First conversation
A 30-minute call to understand the problem, the constraints, and whether an AI system is actually the right shape for it.
30-min intro call - 02 SCOPE
Define the problem space
I map the domain: data sources, edge cases, evaluation criteria. What does a good outcome look like from your users' perspective?
Scope doc · data audit - 03 BUILD
Working prototype
First end-to-end system: model, pipeline, interface. Something you can interact with, demo, and break. Usually two weeks to v0.1.
v0.1 – live and testable - 04 EVAL
Eval and iterate
Benchmarks, failure modes, human evaluation. Numbers honest enough to present to a board. Iterate until the system earns its place.
Benchmark report - 05 SHIP
Production rollout
Integrated with your stack, monitored, documented. The system runs reliably without me watching it.
v1.0 → deployed
Contact
Email is best. I read everything. I reply to most things.
Message received.
Thanks for reaching out. I usually reply within 4-6 hours.